Full disclosure: I am a data scientist not a production-grade SWE.. in any case that sounds about right. I would probably just leave the weights loaded in memory which might be lower latency than querying a db every time you want the score. The DB is mainly there to store the json requests (or however you're receiving them) as well as the predictions for future use.

Another use case for the db would be if someone sends you data to score and you have to append other external data tables to that data, in which case you could use the db to append the data.